TWI839132B - Voice activity detection system - Google Patents

Abstract

A voice activity detection (VAD) system includes a voice frame detector that detects a voice frame during which a voice signal is not silent; a voice detector that detects presence of human speech according to the voice frame; and a threshold update unit that updates an associated threshold for detecting the presence of humane speech according to result of human speech detection by the voice detector.

Description

Voice Activity Detection System

The present invention relates to a voice activity detection (VAD), and more particularly to a voice activity detection system with an adaptive threshold.

Voice activity detection is used to detect or recognize human speech, and is mainly used for voice processing. Voice activity detection can be used to activate voice-based applications. During non-voice periods, voice activity detection shuts down some processing to avoid unnecessary transmission, thereby reducing communication bandwidth and power consumption.

Traditional voice activity detection systems are prone to errors or unreliability, especially in noisy environments. Therefore, it is urgent to propose a novel mechanism to overcome the shortcomings of traditional voice activity detection systems.

In view of the above, one of the purposes of the embodiments of the present invention is to provide a voice activity detection system with an adaptive threshold value, which can be applied to environmental changes and noise overcoming, thereby outputting reliable and correct detection results.

According to an embodiment of the present invention, the voice activity detection system includes a voice frame detector and a voice detector. The voice frame detector is used to detect the voice frame when the voice signal is not static. The voice detector detects human speech based on the voice frame.

In one embodiment, the voice activity detection system further includes a threshold value updating unit, which is used to update the corresponding threshold value to detect human speech according to the result of the voice detector detecting human speech.

100: Voice activity detection system

100A: Voice Activity Detection System

100B: Voice Activity Detection System

11: Converter

12: Voice frame detector

13: Voice detector

14: Critical value update unit

15: Controller

16: Image sensor

17: Artificial Intelligence Engine

18: Speech recognition unit

19: Face recognition unit

200: Voice activity detection method

21: Convert sound into voice signals

22: Detect the voice frame when the voice signal is not static

23: Whether human speech is detected

24: Update critical value

TH_B: First critical value

TH_C: Second critical value

The first figure shows a block diagram of a voice activity detection system of an embodiment of the present invention.

The second figure shows a flow chart of the voice activity detection method of an embodiment of the present invention.

Figure 3A shows the waveforms of voice signals and endpoints.

Figure 3B shows the volume and high-frequency difference of the voice signal.

Figure 3C shows the audio frame.

Figure 4A shows the waveform of the voice signal and endpoints.

Figure 4B illustrates the autocorrelation and the corresponding first critical value TH_B.

Figure 4C illustrates the normalized squared difference and the corresponding second critical value TH_C.

Figure 5A illustrates the autocorrelation and how to obtain the updated first critical value.

Figure 5B illustrates the normalized squared difference and how to obtain the updated second critical value.

Figure 6 shows a block diagram of the voice activity detection system of the first exemplary embodiment of the present invention.

FIG. 7 shows a block diagram of a voice activity detection system of the second exemplary embodiment of the present invention.

The first figure shows a block diagram of a voice activity detection (VAD) system 100 according to an embodiment of the present invention, and the second figure shows a flow chart of a voice activity detection method 200 according to an embodiment of the present invention.

The voice activity detection system 100 of this embodiment may include a transducer 11, such as a microphone, for converting sound into (electronic) voice signals (step 21).

The voice activity detection system 100 may include a voice frame detector 12 that receives a voice signal and is used to detect a voice frame when the voice signal is non-static (step 22). In one embodiment, the voice frame detector 12 uses end-point detection (EPD) to determine the endpoints of the voice signal between which the voice signal is non-static. In one embodiment, the amplitude of the voice signal greater than a preset threshold (which represents the volume) is determined as the endpoint. In another embodiment, the high-order difference (HOD) of the voice signal greater than a preset threshold (which represents the slope) is determined as the endpoint. Figure 3A shows the waveform of the voice signal and its endpoints, Figure 3B shows the volume and high-order difference of the voice signal, and Figure 3C shows the voice frame.

The voice activity detection system 100 of this embodiment may include a voice detector 13 to detect human speech based on the voice frame (step 23).

其中τ代表延遲時間，s代表語音訊框，i=0,…,n-1。 In this embodiment, when the value of the similarity or correlation between speech frames is greater than the corresponding threshold value, the speech detector 13 detects human speech. In this case, an auto-correlation (function) is performed on the speech frame to determine the auto-correlation value, which represents the similarity (or detected pitch) between the speech frame and the (delayed) speech frame with a delay time. The auto-correlation function (ACF) can be expressed as follows:

Where τ represents the delay time, s represents the speech frame, and i=0,…,n-1.

In this embodiment, a normalized squared difference (function) is further performed on the voice frame (eg, a voice frame and a voice frame with a delay time) to determine a normalized squared difference value. The normalized squared difference function (NSDF) can be expressed as follows:

In this embodiment, when the autocorrelation value is greater than the first critical value and the normalized square difference is greater than the second critical value, human speech is detected. The fourth figure A illustrates the waveform of the voice signal and the endpoints, the fourth figure B illustrates the autocorrelation and the corresponding first critical value TH_B, and the fourth figure C illustrates the normalized square difference and the corresponding second critical value TH_C.

Returning to the second figure, if human speech is detected, another voice frame is detected. If human speech is not detected (indicating that noise is detected), before detecting another voice frame, the critical value corresponding to the similarity between voice frames is updated (or adjusted) in step 24. In this way, the voice activity detection system 100 and the voice activity detection method 200 can adapt to determine the critical value based on the human speech detection result, so as to adapt to the current environment, rather than using a fixed critical value like the traditional voice activity detection system and method.

The voice activity detection system 100 of this embodiment may include a threshold value updating unit 14, which is activated by an activation signal (sent by the voice detector 13) to determine the update of the (first/second) threshold value (when no human speech is detected). When no human speech is detected, the activation signal becomes active.

Figure 5A illustrates the autocorrelation and how to obtain the updated first critical value. In this embodiment, the autocorrelation value without delay (i.e., ACF(0)) is subtracted from the maximum autocorrelation value within a specific range (e.g., max(ACF(62:188))) to obtain the updated first critical value.

Figure 5B illustrates the normalized squared error and how to obtain the updated second critical value. In this embodiment, the updated second critical value is equal to the maximum autocorrelation value within a specific range (e.g., max(ACF(62:188))).

According to the above-mentioned embodiment, since the critical value when detecting human speech is adaptively determined, the voice activity detection system 100 and the voice activity detection method 200 can adapt to environmental changes and overcome noise, thereby outputting reliable and correct detection results.

FIG6 shows a block diagram of a voice activity detection system 100A of the first exemplary embodiment of the present invention. In this embodiment, (only) when human speech is detected, the voice detector 13 sends a voice trigger signal to the controller 15, which sends an image trigger signal to wake up the image sensor 16 (such as a contact image sensor (CIS)) to capture an image. It is worth noting that the image sensor 16 is usually in a low power mode or sleep mode until the image trigger signal becomes active. In this way, power consumption and communication bandwidth can be greatly reduced.

In this embodiment, the voice activity detection system 100A may include an artificial intelligence (AI) engine 17, such as a neural network, to analyze the image captured by the image sensor 16 and transmit the analysis result to the controller 15, which executes a specific function or application based on the analysis result.

FIG. 7 shows a block diagram of a voice activity detection system 100B of the second exemplary embodiment of the present invention. The voice activity detection system 100B of FIG. 7 is similar to the voice activity detection system 100A of FIG. 6, and the differences are described as follows.

In this embodiment, the voice activity detection system 100B may further include a voice recognition unit 18, which is used to recognize the spoken language and even translate the spoken language into text according to the voice frame (of the voice frame detector 12), or to recognize the speaker, or to perform both.

The voice activity detection system 100B of this embodiment may further include a face recognition unit 19 for recognizing faces from the image captured by the image sensor 16. The face recognition unit 19 is activated only when the image trigger signal (of the controller 15) becomes active.

The above is only a preferred embodiment of the present invention and is not intended to limit the scope of the patent application of the present invention; any other equivalent changes or modifications that do not deviate from the spirit disclosed by the invention should be included in the scope of the patent application described below.

100: Voice activity detection system 11: Converter 12: Voice frame detector 13: Voice detector 14: Threshold update unit

Claims (11)

A voice activity detection system includes: a voice frame detector for detecting a voice frame when a voice signal is not static; a voice detector for detecting human speech based on the similarity or correlation between voice frames, the voice detector performing autocorrelation on the voice frame to determine an autocorrelation value, which represents the difference between the voice frame and a voice frame with a delay time. similarity; and a critical value updating unit, which is used to update the corresponding critical value to detect human speech according to the result of the speech detector detecting human speech; wherein the speech frame detector uses endpoint detection to determine the endpoints of the speech signal, and the speech signal is not static between the endpoints, and only when human speech is not detected, the critical value updating unit updates the corresponding critical value. The voice activity detection system of claim 1 further comprises: a converter for converting sound into the voice signal. As in claim 1, the voice activity detection system, wherein the amplitude or high-order difference value of the voice signal greater than a preset critical value is determined as an endpoint. As in the voice activity detection system of claim 1, when the similarity value between the voice frames is greater than the corresponding critical value, the voice detector detects human speech. A voice activity detection system as claimed in claim 1, wherein the voice detector performs a normalized square difference on a voice frame and a voice frame with a delay time to determine a normalized square difference value. As in claim 5, the voice activity detection system, wherein when the autocorrelation value is greater than the first critical value and the normalized squared difference is greater than the second critical value, human speech is detected. A voice activity detection system as claimed in claim 6, wherein the autocorrelation value without delay is subtracted from the maximum autocorrelation value within a specific range to obtain an updated first critical value, and the updated second critical value is equal to the maximum autocorrelation value within the specific range. The voice activity detection system of claim 1 further comprises: a controller, which receives a voice trigger signal from the voice detector when human speech is detected; and an image sensor, which sends an image trigger signal to wake up the image sensor in low power mode when human speech is detected. The voice activity detection system of claim 8 further comprises: an artificial intelligence engine for analyzing the image captured by the image sensor and transmitting the analysis result to the controller, which executes a specific function or application based on the analysis result. The voice activity detection system of claim 9 further comprises: a voice recognition unit, which is activated only when the image trigger signal becomes active, to recognize the spoken language or to recognize the speaker. The voice activity detection system of claim 9 further comprises: a face recognition unit, which is activated only when the image trigger signal becomes active, to recognize a face from the image captured by the image sensor.

Images